ASLCitizen-3D

Final SMPL-X motion reconstructions derived from ASL Citizen videos with the MV-Sign extraction pipeline.

Project: https://github.com/ronggong000/MV-Sign

Contents

83,325 final reconstructed SMPL-X motion NPZ files derived from ASLCitizen videos, one per sample. These are predicted motions, not original 3D ground truth. Raw videos, intermediate AiOS/WiLoR predictions, meshes, weights, and SMPL-X/MANO model files are not included. The three official ASLCitizen split CSVs are used to assign each reconstructed sample to its upstream train, validation, or test split. Exact reconstructed split counts are in manifest.json.

Format and Integrity

The tar shards contain <split>/<sample_id>.npz. NPZ files are already compressed; the outer tar container is not compressed. samples.csv records archive members, per-file SHA-256, shape schema identifiers, frame counts, and FPS. manifest.json and SHA256SUMS describe archive-level checksums. array_schema.json lists exact array dtypes and representative shapes. Final NPZ payloads are preserved without recomputing motion parameters.

Load using numpy.load(path, allow_pickle=False). poses contains SMPL-X local axis-angle poses in radians; betas stores fitted shape parameters. Read each file's mocap_framerate, rather than assuming a fixed FPS. Do not interpret trans or cam_trans as calibrated ground-truth camera parameters, or assume saved keypoints were recomputed after every motion repair.

Download with huggingface_hub.snapshot_download(repo_id="mage0001/ASLCitizen-3D", repo_type="dataset", local_dir="ASLCitizen-3D"), then verify and extract:

sha256sum -c SHA256SUMS
mkdir -p extracted
for archive in data/*.tar; do tar -xf "$archive" -C extracted; done

Attribution and Access

Source dataset: ASL Citizen, Desai et al., 2023. Please cite ASL Citizen and MV-Sign, and follow the original dataset terms and applicable SMPL-X/MANO terms. Original RGB data and model assets are not bundled. This resource does not grant additional rights to upstream materials.